Using multilevel models to incorporate spatial heterogeneity in landslide susceptibility assessment

Abstract

Landslide susceptibility is defined as the likelihood of landslide occurrence under favorable terrain conditions. Traditional logistic regression models often fall short by assuming spatial homogeneity, ignoring the inherent variability of physical and environmental factors. To overcome this limitation, multilevel models that incorporate both fixed and random effects to capture variability among different spatial units are used. In this study, a multilevel logistic regression model is employed to assess landslide susceptibility in the Colombian Andes, incorporating both natural hydrological divisions and clustering-based spatial techniques. This approach provides a more nuanced representation of susceptibility by addressing geographic heterogeneity. Two regionalization approaches are tested: natural basin divisions (Atrato, Cauca, Magdalena), and spatial clustering derived from morphometric characteristics. The results reveal that basin-based regionalization achieves a better model fit based on Akaike (AIC) and Bayesian (BIC) information criteria metrics compared to the spatio-clustering model. This suggests that hydrological boundaries are more consistent in capturing landslide dynamics in the study area, likely due to their alignment with geomorphological processes. The fixed and random effect coefficients illustrate that variables such as elevation, relief, and area have significant variability depending on the regionalization method used. A major advantage of using multilevel models is their ability to incorporate spatial variability without artificially dividing the study area and training separate models for each subregion. By using shared information between different spatial units, these models can improve the estimation of both global and local landslide effects.